VI Decision Models in Support of Operational Purchasing Decisions

Chapter V: The available decision models for supplier selection lack differentiation

VThe available decision models for supplier selection lack differentiation

The previous chapter has made clear that supporting supplier selection requires a differentiated collection of decision models. In this chapter, we present the results of an extensive investigation of the literature regarding the available decision models for supplier selection. The position of this chapter in the overall step-wise planning is shown in figure 5.1.

Figure 5.1: Positioning of chapter V

The purpose is to find out to which extent the required differentiation is present in the collection of available decision models. The results suggest that although a variety of techniques are used, the existing body of literature on decision models for supplier selection suffers from one-sidedness in its focus and assumptions.

Purchasing and OR researchers have focused on one-dimensional and compensatory choice models

In this section, we present the results of an extensive literature search on decision models for supplier selection. The search included both textbooks and journals. The vast majority of the models found apply to the choice phase of the supplier selection process. Furthermore, the decision rule used is often such that a low score on one criterion can be compensated by a high score on another criterion.

Categorical models and Neural Networks are choice models with implicit decision rules

Intuitive explanation of categorical models

Basically, categorical models (see Stevens, 1978; Zenz, 1981; Lee and Dobler, 1977) are qualitative models. Based on historical data and experience current suppliers are evaluated on a set of criteria. These evaluations consist of qualifying the performance of suppliers with respect to a criterion either as 'positive', 'negative' or 'neutral'. After the supplier has been rated with respect to all the criteria, the evaluator gives an overall rating, again through ticking one of the three options.

An example of a categorical model

Based upon Dobler and Burt (1996) we present the following typical example of a categorical method (see figure 5.2).

Figure 5.2: Example of a categorical method template (Based on Dobler and Burt, 1996)

The representatives of the various departments that participate in the selection process, rate the supplier with respect to the criteria through ticking either ‘preferred’, ‘neutral’ or ‘satisfactory’ for each criterion. After the supplier has been rated with respect to all criteria, each department gives an overall rating.

Categorical models are flexible models but they do not formalise the ultimate choice

Clearly, in a categorical model, both quantitative and qualitative criteria may be used. In addition, the model is easy to use and to understand. The flexibility of the categorical method implies that the method can be used in many situations. However, the ultimate choice between suppliers is not formalised, i.e. the aggregation of the scores on the various criteria into the overall score is not obvious: trade-off’s and decision rules remain implicit.

Intuitive explanation of a Neural Network model for supplier selection

Albino and Garavelli (1998) propose the use of a Neural Network approach to supplier selection. Neural Networks are systems that can be said to ‘pick up’ or learn the decision rule that implicitly is used in a series of different decision situations which are fed through the Neural Network. Applied to supplier selection, it means that first the scores on the criteria as well as the overall-score (of a sufficiently big number of suppliers and bids) are ‘shown’ to the Neural Network. From these (learning) cases, the Neural Network extracts the underlying decision rule, i.e. it derives a rule for converting criteria scores into an overall rating. In case of a new (yet comparable) supplier selection, the characteristics of the situation (i.e. the scores of the suppliers on the criteria) are fed into the Neural Network model. Next, the model uses the decision rule it has learned to arrive at the overall rating of the new suppliers.

An example of a Neural Network model applied to supplier selection

In their article, Albino and Garavelli discuss the application of a Neural Network model in rating subcontractor bids in the construction business. First, a so-called training set was established, see table 5.2

Table 5.2: Training set for building a Neural Network supplier selection model (based on Albino and Garavelli, 1998)

A purchaser evaluated the bids by expressing the overall preference for a bid in a number between 1 and 10. Based on the training set, the Neural Network model was constructed. This model was then tested on 5 new bids (see table 5.3).

Table 5.3: Testing of the Neural Network model to 5 new bids (Albino and Garavelli, 1998)

It is clear that the Neural Network model very closely ‘follows’ the overall ratings by the purchaser.

Neural Networks capture imprecision but are restricted to repetitive situations

The authors claim that the Neural Network is especially useful in ill-structured situations, which are highly affected by uncertainty and imprecision. Indeed, the approach ‘frees’ the purchaser from having to precisely explicate his mental processes. But similar to the categorical method, the decision rule remains ‘invisible’. The advantage of the Neural Network (compared to the categorical method) is that it assures a consistent execution of the implicit decision rule. Evaluation tableau’s from the categorical method, such as depicted in table 5.1 could be used as input for building Neural Network models. However, just as with the categorical method, not having an explicit decision rule (structure) turns into a disadvantage when it comes to justifying and explaining the supplier selection to others. The argument: “We chose this supplier because it fits best with our previous selection-behaviour” may not be convincing or acceptable for other suppliers or senior management. Furthermore, as the authors indicate, the learning phase of a Neural Network is based on a specific (recurring!) supplier selection situation. When this situation changes, the network has to be trained again.

Cost-ratio, TCO and Decision Analysis models strictly include quantitative criteria

Intuitive explanation of cost-based models

Cost-ratio models (see Stevens, 1978; Zenz, 1981; Lee and Dobler, 1977) are quantitative, cost-based models in which (procurement) costs are associated with a particular supplier.

More recently, the cost ratio logic has been extended to what usually is referred to as Total Cost of Ownership (TCO) models (see e.g. Ellram, 1994). Compared to costratio methods, TCO-models usually cover additional costfactors as well as a more sophisticated treatment of data, e.g. TCO- models are often more advanced in terms of computerised support and the use of various analytical techniques such as forecasting and price analysis.

Some examples of costbased models

For a particular supplier, the costs related to inspection, delivery and service are expressed as a percentage of the total value of goods that are purchased from that supplier. These percentages (cost-ratios) are used again when evaluating new bids from these suppliers. The following example (see table 5.4) further illustrates this.

Table 5.4: Example of cost-ratio method

A typical example of a TCO-model is presented in Smytka and Clemens (1993):

Table 5.5: Example of TCO-model (based on Smytka and Clemens, 1993)

Similar to the costratio method, the prices stated in the suppliers’ quotations are the basis for the calculation of what is considered a fair estimate of the total costs of doing business with those suppliers.

An example showing such calculations is given in table 5.6.

Table 5.6: Example of TCO calculations (based on Smytka and Clemens, 1993)

Costbased models have a practical appeal but may be costly to operate and maintain

Cost-ratio as well as TCO-models ultimately express the performance of suppliers in clear, comparable monetary units. This definitely gives these models a practical appeal. However, it is also clear that practical use of such models requires several conditions to be met.

First, the cost information concerning suppliers must be collected and stored. Obviously, the benefits of maintaining and operating such an information system must outweigh the costs.

Secondly, it seems that in case of completely new suppliers and products it may not be possible to obtain the required information. Finally, the costbased models do not facilitate the evaluation of other (qualitative) criteria when selecting suppliers.

An intuitive explanation of Decision Analysis

In Decision Analysis (DA) applications, the costs of choosing a supplier are not assumed fixed and given but variable and dependent on the situation. DA aids the purchaser in choosing the supplier that yields the lowest expected costs.

A formal notation of a Decision Analysis model for supplier selection

In formal terms, a DA model for supplier selection may look as follows:

Where:

S* = the supplier that yields the lowest expected costs;

S = the set of suppliers considered;

N = the number of possible situations;

Cij = the costs related to supplier j in situation i;

Pi = the probability that situation i will occur.

An example of a DA- model for supplier selection

Soukup (1987) describes a clear example of DA applied in a supplier selection setting. In this particular example, the uncertainty applies to the demand for the item purchased. Naturally, other relevant factors, e.g. price or delivery performance, may also be uncertain. The data used in the example are summarised in table 5.7.

Table 5.7: Example of DA-model for supplier selection (based on Soukop, 1987)

In this example, supplier C is preferred over the other suppliers even though this supplier quotes the highest prices for smaller order sizes. Intuitively, we might therefore want to order from supplier A and /or supplier B, but the Decision Analysis approach points out the significant cost advantage of supplier C in case of high demand.

Decision analysis covers (quantitative) uncertainty but is restricted to repeating situations

A distinctive advantage of DA is that it explicitly deals with the ever-present uncertainty about the value of such important purchasing variables as price, demand and supply quantities. In that respect, many of the other decision models for supplier selection are deterministic. However, it is also clear that DA does not allow for the simultaneous evaluation of other (qualitative) criteria. In addition, the results of a DA-model require a proper and cautious interpretation. The ‘expected’ costs are the average of costs that occur in a sequence of comparable purchasing situations. In other words: the result has little meaning in one-off, unique purchasing situations.

Linear Weighting and Weighted Product models employ compensatory decision rules

An intuitive explanation of linear weighting models

A large number of models originate from the linear weighting principle (also often referred to as weighted point plans). Examples of the general version (a) can be found in Baily and Farmer (1990), Stevens (1978), Zenz (1981) and Lee and Dobler (1977). The principal idea is to assign numerical scores to a set of quantitative and/or qualitative criteria, express the relative importance of every criterion in numerical terms and subsequently determine the composite performance index by multiplying each score with its numerical weight and adding up all the resulting products.

A formal notation of a linear weighting model for supplier selection

In more formal terms, the basic linear weighting model for supplier selection can be defined as below:

Where:

pi = overall preference for supplier i;

n = number of criteria;

sij = score of supplier I on criterion j;

wj = weight of criterion j.

In the following subsections we further elaborate on the elements of this basic model.

An example of a linear weighting model for supplier selection

Pinkerton (1986) provides a typical example of the basic linear weighting model. In this example, it is assumed that we have received three bids for a particular item (see table 5.8).

Table 5.8: Available information on suppliers and their bids

In addition to the available information on the prices, we also have information concerning the suppliers’ performance with respect to quality and delivery. In table 5.8 the quality and delivery ratings are expressed as the percentage of error-free items received in the past from a supplier and the percentage of items delivered in time in the past by a supplier, respectively. As the latter two figures are percentages, their maximum value is 100. In order to make price comparable to delivery and quality, Pinkerton suggests to assign a rating of 100 to the lowest bid, i.e. supplier A, and calculate the ratings of supplier B and supplier C as follows:

rating supplier B = (656500)/(706438) = 93%

rating supplier C = (656500)/(667750) = 98%

Next, a weight is assigned to each criterion. Now the final composite performance of the suppliers can be calculated, as shown in table 5.9

Table 5.9: Example of linear weighting model calculations (Pinkerton, 1986)

In this example, using the linear weighting model, supplier A would be chosen.

There are several ways of deriving weights for the linear weighting model

In most discussions on linear weighting models, it is assumed that the purchaser is able (or feels comfortable with) directly assigning a point estimate value (between 0 and 1) to each criterion in such a way that together the weights sum up to 1. However, several authors propose more sophisticated techniques.

Williams (1984) proposes the use of conjoint analysis in deriving the weights of the criteria. Basically, there are two ways of doing this. The first approach consists of asking the purchaser to rank various (hypothetical) combinations of scores on all criteria and subsequently infer the best fitting weights from these evaluations, e.g. by means of regression analysis. Alternatively, in a two-at-a-time approach, the purchaser ranks various levels of two criteria at a time. After that again, the best fitting weights are inferred using a regression program. Obviously, using conjoint analysis, the purchaser does not have to directly assign weights to the criteria. In that respect, uncertainty and imprecision are accommodated. However, in some way the problem of directly assigning weights is transported to the problem of directly assigning overall scores to different sets of scores on the criteria. In addition, for many purchasers regression techniques may not be straightforward and for the method to be practical, software should be available to perform the calculations.

Min (1994) applies the so-called ‘indifference-trade off’’ method in order to derive weights within a Multi-Attribute-Utility Theory (MAUT) framework. MAUT will be discussed in more detail further on. The indifference trade-off method bears much resemblance to the conjoint analysis method. In MAUT, two suppliers which have different scores but in the trade-off are equally preferred, must have the same overall scores. Thus, the purchaser is asked to specify different sets of criteria scores that are all equally preferred. From these sets, together with the constraint that the weights must sum up to 1, the weights can be calculated. As such, this method (also) frees the purchaser from having to directly assign numerical weights to the criteria. However, again one might argue that the problem of direct assignment of numerical scores is not solved but rather shifted to another phase in the process, which in this case concerns the construction of value functions in the MAUT method. These value functions translate the raw scores on the criteria into normalised scores between 0 and 1.